Mistral Large 4: the weights are here, the cloud challenge begins

This time, there is something to download

On 6 October, Mistral introduced Large 4 as an API preview and promised its weights by the end of the month. As of 29 October 2026, the official repository contains the model files: roughly 1.06 TB, split across 54 Safetensors files. Moving from a promise to an available artifact is the real news.

A large model, more than another download

The technical documentation describes a Mixture-of-Experts architecture with approximately 1.05 trillion total parameters and 52 billion active parameters, accepting text and image inputs. Only part of the model participates in each computation step; that does not remove the memory requirements.

An NVFP4 variant is also available for Blackwell GPUs. The vendor’s performance claims still need checking against your own tasks. We are most interested in what can run reliably.

Open weights, with specific conditions

One detail matters: the Mistral Ecosystem Open-Weight License is not simply an unrestricted MIT license. The license text requires a separate commercial agreement for a Model-as-a-Service business. It distinguishes this activity from internal use and certain capabilities embedded in a product.

For a hosting provider, that distinction affects how an offering is prepared. Being able to download a model does not establish unrestricted permission to resell access to it.

The interesting cloud work comes afterwards

Choosing an inference engine, measuring time to first token, handling load and isolating customers determine the actual experience. One successful demo response cannot tell you how many concurrent requests a platform can handle, or what each request will cost.

For automation, we would start with concrete tasks: explaining an error log, suggesting a fix or preparing a configuration. Then we would test the boundaries: tool permissions, human approval and the ability to undo an action. A capable model deserves an equally sound execution environment.

At Novoid, this release informs our research into AI and infrastructure. It is not an announcement that Large 4 has been deployed in our services. The opportunity is to compare, experiment and make decisions based on measured results rather than a leaderboard.

Mistral’s original announcement. Conceptual illustration generated for this article.